Professional Documents
Culture Documents
Tuberculosis disease remains a global threat and a leading cause of death. The world
health organization stressed on increasing the decline rate of the disease. Early diagnosis
and evaluation of the TB severity stage are important for determining the right treatment
and eventually avoiding the death of curable patients cases. The voluminous amount of
medical images available and the proven performance of deep learning in medical diagnosis
are a motive for an automatic medical diagnosis to tackle the high demand for radiologists
and reduce the costs of diagnosis. Furthermore, there has been a considerable growth in
recent years in the field of deep learning, which allows performing hard classification
tasks. Particularly in computer vision fields, it has been tested and proven that deep
convolutional neural networks (CNN) are very promising algorithms for various visual
tasks. Thus, in this report, we are interested in an automatic TB severity scoring by
applying deep learning techniques on annotated chest CT scans. In ImageCLEF2018
Tuberculosis Severity Score task, the results obtained were very good, supporting the
applicability of assessing the severity based only on the CT image. Furthermore, explore
various ways to avert the weak points of the work of the state of the art. Eventually,
reaching better results.
1
Contents
Introduction 4
General Conclusion 22
2
Abbreviations
3D 3 Dimensions
BK Bacillus Koch
CT Computed Tomography
FN False Negatives
FP False Positives
ML Machine Learning
TB Tuberculosis
TN True Negatives
TP True Positives
3
Introduction
In recent decades, medical imaging has become indispensable in the diagnosis and
therapy of diseases. With the enhancement of medical imaging databases, new methods
are required to better handle this huge volume of data. However, because of the large
variations and complexity of medical imaging data, it is generally difficult to deduce
analytical solutions or simple methods to describe and represent objects such as lesions and
anatomies in data. Therefore, medical imaging tasks require learning from the examples,
and this is one of the key interests of the machine learning field.
Machine learning has become one of the main tools for medical image analysis. Machine
learning techniques are solutions for developing tools to help physicians diagnose, predict,
and prevent the risk of disease before it becomes too late in less time. Deep Learning
is a new component in the field of machine learning that encompasses a wide range of
network architectures designed to perform multiple tasks. The first use of neural networks
for medical image analysis goes back more than twenty years, their use has increased by
several orders of magnitude over the last five years. different, articles [1, 2, 3, 4, 5] have
highlighted the application of deep learning to a wide range of medical image analysis tasks
(segmentation, classification, detection, recording, image reconstruction, enhancement,
etc.).
The goal of our project is to automatically give score of TB severity via Computed
Tomography (CT) scan. One of the possible applications of this study is to accelerate the
diagnosis of the disease from a radiology image without resorting to expensive medical
tests. This work is part of the ImageCLEF2018 task of classifying types of TB, which has
shown more promising results. We explore in this paper the different work and different
concepts that link with this problem. Starting by giving an overview of tuberculosis and
its types in chapter 1. Then, discuss the relationship between artificial intelligence and
4
medical images then give definition of some important parts of these fields in chapter 2.
In chapter 3, the ImageCLEF Tasks are described and the related work of tuberculosis
severity scoring is reviewed.
5
Chapter 1
Tuberculosis can develop rapidly after the first contact with the microbe, but it can
also appear several years later.
Globally, the best estimate is that 10.0 million people (range, 9.0–11.1 million) devel-
oped TB disease in 2017: 5.8 million men, 3.2 million women and 1.0 million children
[11]. There were cases in all countries and age groups, but overall 90% were adults (aged
15 years), 9% were people living with Human Immunodeficiency Virus (HIV) (72% in
Africa) and two thirds were in eight countries: India (27%), China (9%), Indonesia (8%),
the Philippines (6%), Pakistan (5%), Nigeria (4%), Bangladesh (4%) and South Africa
(3%). These and 22 other countries in World Health Organization (WHO)’s list of 30
high TB burden countries accounted for 87% of the world’s cases[11]. Only 6% of global
6
cases were in the WHO, European Region (3%) and WHO Region of the Americas (3%)
[11].
7
Estimated TB incidence rates are shown in Figure 1.1.
Focal TB Focal TB is diagnosed in the event of lesions not exceeding two lung segments.
A small number of pathological foci with a diameter of about 1 cm are formed in the lung,
which occurs at different times.[17]
8
Infiltrative Tuberculosis infiltrative tuberculosis is characterized by the fuzzy out-
lines of shadows on radiography. This often happens when forming new cavities. infiltra-
tion appears on radiography as rounded, or shade of homogeneous structure. All spots
are divided into small, medium and large. The small ones have a size of 1 to 2 cm, the
means 2-4 cm, the large ones up to 6 cm.[18]
Summary
As shown in this chapter, tuberculosis is a chronic and may result in the death of the
infected person. Furthermore, TB has several types that are diagnosed and treated dif-
ferently. In the next chapter, the different medical imaging technologies used to diagnose
9
TB are presented along with the role of machine learning to analyse medical imaging. A
particular focus will be put on deep learning and its architectures and different evaluation
metrics.
10
Chapter 2
X-ray imaging this technology is the oldest and the most frequently used, it works on
wavelengths and frequencies that can penetrate through the skin creating a visualization
of the inner body. It used to detect skeletal system malfunctioning, cancer through
mammography, and other diagnoses that involve the visualization of the inner body. This
technique comes with risks associated with the use of X-ray radiation.
Ultrasound imaging uses High-frequency sound waves that are transmitted from the
probe to the body via the conducting gel, those waves then bounce back when they hit
the different structures within the body and that is used to create an image for diagnosis.
This technology is considered the safes without any recorded side effects of its usage and
is the most cost-effective. Due to its low risk, it is the first choice for pregnancy.
Magnetic resonance imaging uses a strong magnetic field and radio waves it enables
an in-depth view of the inside of a joint or ligament to be seen, rather than just the
outside as in the case of CT scans and X-ray. It has risks associated with the use of a
11
strong magnetic field where any kind of metal implant, artificial joint could be moved or
heated up within the magnetic field.
There exist several medical images file formats all of them sharing the goal of standard-
izing medical images storage and transmission. Major file formats widely used in medical
imaging are Analyze, Neuroimaging Informatics Technology Initiative (Nifti), Minc, and
Digital Imaging and Communications in Medicine (Dicom).
Analyze Analysis 7.5 was created in the late 1980s as a format used by the Analyze
commercial software developed at the Mayo Clinic in Rochester, MN, USA. For more than
a decade, the format was the standard for post-processing medical imaging. The big point
of view of the Analyze format is that it was designed for multidimensional (volume) data.
Indeed, it is possible to store 3D or 4D data in a file (the fourth dimension being typically
the temporal information). An Analyze 7.5 volume includes two binary files: an image
file with the extension .img which contains the raw voxel data, and a header file with the
extension .hdr which contains the metadata (number of pixels in the three dimensions,
voxel size, and data type). The header has a fixed size of 348 bytes and is written as a
structure in C programming language. Reading, and editing the header requires a utility
software. The format is now considered ”old” but it is still widely used and supported by
many processing software, viewers, devices, and conversion utilities.[22]
Nifti Nifti is a file format created in the early 2000s with the intention to create a format
that preserves compatibility of the Analyze format but solving its weaknesses. Nifti may
be considered a revised ’Analyze’ format. NIfTI uses the ”empty space” in the ANALYZE
7.5 header to add several new features such as image orientation with the intention of
avoiding left-right ambiguity in the brain study. In addition, Nifti includes unsupported
data type in the Analyze format as an unsigned 16-bit format. Although the format also
allows the storage of header and pixel data in separate files, the images are usually saved
as a single ’.nii’ file in which the header and data are stored. pixels are merged. The
header has a size of 348 bytes in the case of data storage ’.hdr’ and ’.img’ and a size of
352 bytes in the case of a single file ’.nii’. This difference in size is due to the presence of
four additional bytes at the end, essentially to make the size a multiple of 16, and also to
provide space for storing additional metadata.[23, 22]
Minc The Minc file format was developed in 1992 to provide a flexible data format
for medical imaging. The first version of the Minc format (Minc1) was based on the
standard common network format (NetCDF). Subsequently, to overcome the large data
file support constraint and provide new features. Minc’s development team chose to
12
upgrade from NetCDF to Hierarchical Data Format version 5 (HDF5). This new version
which is not compatible with the previous one was called Minc2.[24, 22]
Thanks to the rise of powerful and affordable computing power and the appearance
of big data. machine learning is witnessing an exciting evolution. Nowadays, machine
13
learning techniques are being used in almost every field of research and application ranging
from healthcare, medicine, agriculture, and finance . . . etc. As a result of these advances,
systems which only a few years ago performed at noticeably below-human levels can
now outperform humans at some specific tasks and complex games such as chess and
Shogi[27, 28].
Machine learning algorithms are organized in categories based on the type of learning
and whether data is available, labeled or not. Thus, we distinguish supervised learning,
semi-supervised learning, unsupervised learning, Reinforcement learning algorithms.
There exist numerous supervised learning algorithms each has its strengths and weak-
nesses. for now, there isn’t an algorithm that can perform best in every situation[31].
Support vectors machine (SVM) and decision trees are well-known examples of super-
vised learning algorithms and will be discussed next.
14
Decision Trees
A decision tree represents a function that takes as input a vector of attribute values
and returns a “decision”—a single output value[29]. It is one of the predictive modeling
approaches used in statistics, data mining and machine learning. It works as decision
support tool. A decision tree is a flowchart-like structure, works by recursively splitting
training data into subsets based on the value of a single attribute. Each split corresponds
to a node in the tree representing a test on the attribute/features, each branch represents
the outcome of the test, and each leaf node represents a class label (decision taken after
computing all attributes). A basic structure of a decision tree is shown in Figure 2.2.[41,
42] There exist different algorithms for building a decision tree, namely ID3 (Iterative
Dichotomiser 3), C4.5 (successor of ID3), CART (Classification And Regression Tree),
CHAID (CHi-squared Automatic Interaction Detector) and MARS. The algorithms and
techniques of building decision trees are an exciting research domain, where the speed,
efficiency, and accuracy are improved [42]. Decision trees are used in many fields including
medecine[43, 44], astronomy[45], and genetics[46]. . . etc.
Figure 2.3 shows a common architecture of ensemble learning. There exist two types of
ensembles, those that use a single base learning algorithm to produce homogeneous base
learners, leading to homogeneous ensembles, i.e, an ensemble of decision trees. The second
type consists of the use of different base learning algorithm to produce heterogeneous
base learners, leading to heterogeneous ensembles, i.e, an ensemble of neural networks
and SVMs.
15
Figure 2.3 – A common ensemble architecture.
The use of an ensemble leads often to a much stronger generalization than the use of
a single learner. Ensemble methods are proven to boost weak learners and turn them to
strong learners with more accuracy.[48]
Figure 2.4 – A simplified illustration of Hansen and Salamon [1990]’s observation: Ensemble is often
better than the best single.
Boosting and bagging are methods of ensemble learning which are used to improve the
stability and accuracy of base learners. They’re briefly described below.
Bagging stands for bootstrap aggregation which is a method for generating multiple
versions of a predictor and using these to get an aggregated predictor[49]. Random forest
is a well-known example of an ensemble learning algorithm. It uses the combination of
bagging technique[50] and random selection of features on the decision tree as the base
learner to produce a collection of decision trees thus constructing a random decision forest.
16
learning or hierarchical learning. The hierarchy of nature enables the computer to learn
complicated concepts by building them out of simpler ones through the use of layers. the
graphical representation of these concepts shows deep graphs, with many layers. That is
why it’s called deep learning.
There exist many architectures involving deep learning such as deep neural networks,
deep belief networks, deep Boltzmann machines and recurrent neural networks which have
been applied in a wide range of domains, like Computer vision, natural language process-
ing, education, finance . . . etc. due to the vast domain of deep learning, A particular focus
is given to deep neural networks architecture in this paper.[51]
Deep neural networks have shown a state of the art performance that has beaten many
ML algorithms, especially in computer vision. Deep networks could be categorized into
four major network architectures groups: Unsupervised Pretrained Networks (UPNs),
Convolutional Neural Networks (CNNs), Recurrent Neural Networks, Recursive Neural
Networks[52]. Widely used examples of CNN architectures are GoogleNet[53], ResNet[54],
VGGNet[55], AlexNet[56].
Confusion Matrix A confusion matrix is a table that is often used to describe the
performance of a classification model (or ”classifier”) over a set of test data for which the
true values are known. In the case of a binary classifier, we get Table 2.2.
• true positives (TP): These are cases in which the predicted class is 1, and the actual
class is 1.
• true negatives (TN): These are cases in which the predicted class is 0, and the actual
class is 0.
17
Predicted
1 0
1 TP FN
Actual
0 FP TN
Table 2.2 – A confusion matrix of a binary classification.
• false positives (FP): These are cases in which the predicted class is 1, and the actual
class is 0.
• false negatives (FN): These are cases in which the predicted class is 0, and the actual
class is 1.
Root Mean Square Error (RMSE) Root Mean Square Error (RMSE) is the standard
deviation of the residuals (prediction errors). Residuals are a measure of how far from
the regression line data points are; RMSE is a measure of how spread out these residuals
are. In other words, it tells you how concentrated the data is around the line of best fit.
Root mean square error is commonly used in regression analysis to verify experimental
results and machine learning models. Its formula is:
"N #1/2
2
X
RMSEf o = (zfi − zoi ) /N (2.2)
i=1
ROC curve The ROC curve is created by plotting the true positive rate (TPR) against
the false positive rate (FPR) which are calculated from the output of confusion matrix
at various threshold settings. The true-positive rate is also known as sensitivity, recall or
probability of detection in machine learning. The false-positive rate is also known as the
fall-out or probability of false alarm and can be calculated as (1 specificity).
2.3.1 Niftynet
NiftyNet is a deep learning platform for medical imaging built on top of The TensorFlow
framework[57]. It comprises several modular components that eases and abstracts the
model building pipeline. The NiftyNet works by connecting four components: a Reader to
load data from files, a Sampler to generate appropriate samples for processing, a Network
to process the inputs, and an output handler (comprising the Loss and Optimizer during
training and an Aggregator during inference and evaluation). These components are
briefly depicted in Figure2.5.[58]
18
Figure 2.5 – A brief overview of NiftyNet components[58].
Summary
In this chapter, the different medical imaging technologies used to diagnose TB were
presented along with the role of machine learning in medical imaging analysis. A partic-
ular focus was put on deep neural networks and its architectures. In the next chapter,
the severity scoring subtask organized by the evaluation campaign ImageCLEF will be
presented along with the oragnizer. Moreover, the dataset, evaluation metrics used will
be looked at with f the top 3 and MostaganemFSEI teams approaches to automatic tu-
berculosis severity scoring.
19
Chapter 3
This task was proposed by ImageCLEF 2019 which is an evaluation campaign that
is being organized as part of the CLEF initiative labs. SVR task lays under the Image-
CLEFmed Tuberculosis category of tasks along with CT report task (CTR) which consists
of generating an automatic report based on the CT image of a patient.
This task possesses different difficulties and challenges. The small amount of dataset
and examples for the use of deep learning, the large dimensions of the image slices without
the availability of an expert to assist in reducing the irrelevant slices and decreasing the
dimensions. The lack of powerful computation resources makes it difficult to train a
machine learning model, particularly a deep neural network.
In the next sections, we will describe the dataset being used for ImageCLEF 2019
SVR task and discuss the participating teams’ approaches and the results obtained for
the severity scoring of ImageCLEF 2018.
3.1 Dataset
The dataset contains 335 chest 3D CT images with an image size per slice of 512*512
pixels and number of slices varying from about 50 to 400 of TB patients. All the CT
images are stored in NIFTI file format with .nii.gz file extension (g-zipped .nii files). In
addition to the CT images, a set of clinically relevant metadata was given. The selected
metadata includes the following binary measures: disability, relapse, symptoms of TB,
comorbidity, bacillary, drug resistance, higher education, ex-prisoner, alcoholic, smoking,
severity. Moreover, automatically extracted lung masks where provided along with a set
of tools for loading/saving niftii files. 218 patients are used for training and 117 for test.
20
3.2 Participation and results
In 2018 there were 7 teams that participated and submitted their result in the SVR
task. Each team used its own approach to build a solution for the SVR task. the partic-
ipants were allowed to submit up to 10 runs for the task. A run consists of submission of
the results of the developed model when applied on a test set provided by imageCLEF.
The runs submitted for the severity scoring subtask were evaluated in two ways. One
used the original severity scores from 1 to 5 and the task was to predict those numeri-
cal scores as precise as possible (a regression problem). Here, Root Mean Square Error
(RMSE) was computed between the ground truth severity and the predicted scores pro-
vided by the participants. Alternatively, the original severity score was transformed into
two classes, where scores from 1 to 3 corresponded to ”high severity” and the 4 and 5
scores corresponded to the ”low severity” class. In this case, the participants had to pro-
vide the probability of TB cases to belong to the ”high severity” class. The corresponding
results were evaluated using AUC. in this section we present the 3 top performing ap-
proaches in terms of Root Mean Squared Error (RMSE) along with the approach used by
MostaganemFSEI team.
Top RMSE achieved was 0.7840, by the team UIIP BioMed with a single run. A Coder-
Decoder Convolutional Neural Network trained on a third-party dataset of 149 CT scans
with lesions labeled by a qualified radiologist was used as a Lesion-based TB-descriptor
extraction method. The lesion-based TB-descriptors extracted from the patients CT
scans were used to generate a random forest of 100 trees as a classifier[59]. This approach
had the best RMSE 0.7840 and an AUC of 0.7708. The second best RMSE was 0.8513,
achieved by the team MedGIFT among 9 runs. A graph model of the lung based on
regional 3D texture features was used in an SVM classifier [60]. This method resulted
in a better AUC than the previous approach. VISTA@UEvora team ranked 3rd with an
RMSE of 0.8883 among 7 runs. The team used a 3D modeling and further extraction of
texture patterns approach with a multi-layer perceptron algorithm[61]. Finally, the team
MostaganemFSEI used an approach that consists of extracting a single image semantic
descriptor for each CT scan/patient instead of considering all the slices as separate samples
[62]. The semantic descriptor was then passed as an input to bagging of a set of Random
forest learners to perform a hierarchical classification of the severity score. With this
approach, the team achieved an RMSE of 0.9721 and an AUC of 0.5987.
RMSE AUC
Group Name Run RMSE Rank AUC Rank
UIIP BioMed[59] SVR run TBdescs2 zparts3 thrprob50 rf100.csv 0.7840 1 0.7025 6
MedGIFT[60] SVR HOG std euclidean TST.csv 0.8513 2 0.7162 5
VISTA@UEvora[61] SVR-Run-07-Mohan-MLP-6FTT100.txt 0.8883 3 0.6239 21
MostaganemFSEI[62] SVR mostaganemFSEI run3.txt 0.9721 12 0.5987 25
Table 3.1 – Results obtained by the top 3 participants and MostaganemFSEI team in the SVR
subtask [8].
21
General Conclusion
This report presents the results of the studies we conducted to obtain the knowledge
needed to achieve our goal of automatically calculating the tuberculosis severity score via
computed tomography (CT) image analysis.
Through this bibliographic study, we have been able to make several conclusions. First,
Tuberculosis disease continues to lead the list of mortal disease, so there is a need for ways
to increase the rate of decline of the disease. Eventually, automated dignosis and scoring
of tuberculosis severity woud increase the availability and the time of diagnosis. Secondly,
deep learning has imposed its supremacy and proven its efficiency in the field of image
analysis. It is therefore worth the eandeavor to accelerate and improve image analysis.
Finally, the work that has been done by the top performing teams to score pulmonary
tuberculosis severity as part of the international ImageCLEF challenge showed the results
obtained were very promising and yet there is still room for improvement.
In the practical part of this project we will try to apply deep learning and use the
medical deep learning framework niftyNet to explore various ways to avert the weak
points of the work of the state of the art. Eventually, reaching better results.
22
Abbreviations
3D 3 Dimensions
BK Bacillus Koch
CT Computed Tomography
FN False Negatives
FP False Positives
ML Machine Learning
TB Tuberculosis
TN True Negatives
TP True Positives
23
List of Figures
24
List of Tables
25
Bibliography
[2] H. A, L. SCB, L. JS, F. MT, and M. SK, “A shift-invariant neural network for the
lung field segmentation in chest radiography,” A Shift-Invariant Neural Network for
the Lung Field Segmentation in Chest Radiography, vol. 18, no. 3, pp. 241–250, 1998.
[4] K. J. E, N. FD, J. TK, and K. DL, “Abdominal organ segmentation using texture
transforms and a hopfield neural network,” IEEE Transactions on Medical Imaging,
vol. 18, no. 7, pp. 640–648, 1999.
[5] R. WE, G. JO, C. EN, and et al, “Automated segmentation and classification of
multispectral magnetic resonance images of brain using artificial neural networks,”
IEEE Trans Med Imaging, vol. 16, no. 6, 1997.
26
[11] W. H. Organization, “Global tuberculosis 2018 report,” 2018. http://apps.who.
int/iris/bitstream/handle/10665/274453/9789241565646-eng.pdf.
[12] U. N. L. of Medicine, “Pulmonary tuberculosis.” https://medlineplus.gov/ency/
article/000077.htm.
[13] M. A, B. M, T. F, O. R, O. R, and S. H. et al, “Miliary tuberculosis: clinical
manifestations, diagnosis and outcome in 38 adults,” Respirology, vol. 6, p. 217–24,
Sept. 2001.
[14] S. SK, M. A, S. A, and M. DK, “Miliary tuberculosis: new insights into an old
disease,” Lancet Infect Dis, vol. 5, pp. 415–30, July 2005.
[15] B. HM, B. WJ, C. RE, D. CL, E. SC, and F. L. et al, “American thoracic soci-
ety/centers for disease control and prevention/infectious diseases society of america:
treatment of tuberculosis,” Am J Respir Crit Care Med, vol. 6, pp. 603–62, Feb.
2003.
[16] R. L. Hunter, “Pathology of post primary tuberculosis of the lung: an illustrated
critical review.,” Tuberculosis, vol. 91 6, pp. 16–17, 2011.
[17] M. Format, “Focal pulmonary tuberculosis - causes, symp-
toms, diagnosis and treatment.” http://medicalformat.com/
1194-focal-pulmonary-tuberculosis-causes-symptoms-diagnosis-and-treatment.
html.
[18] Respiratory, “infiltrative tuberculosis - causes, symptoms, diagnosis
and treatment.” http://respiratory.vsebolezni.com/tuberkulez/
infiltrativnyj-tuberkulez-legkix-chto-eto-takoe.html.
[19] S. D. Pitlik, V. Fainstein, and G. P. Bodey, “Tuberculosis mimicking cancer 2014;a
reminder,” The American Journal of Medicine, vol. 76, pp. 822–825, May 1984.
[20] S. Vento and M. Lanzafame, “Tuberculosis and cancer: a complex and dangerous
liaison,” The Lancet Oncology, vol. 12, pp. 520–522, jun 2011.
[21] S. Jeffrey, H. Shen, K. Vadim, and J. Chinnawsamy, “Pathogenesis of post primary
tuberculosis: immunity and hypersensitivity in the development of cavities,” Annals
of clinical and laboratory science, vol. 44, pp. 365–87, jun 2015.
[22] M. Larobina and L. Murino, “Medical image file formats,” Journal of Digital Imaging,
vol. 27, pp. 200–206, Dec 2013.
[23] M. Jenkinson, “Nifti-1 data format,” Oct 2007. https://nifti.nimh.nih.gov/
nifti-1.
[24] R. D. Vincent, L. Baghdadi, and V. S. FONOV, “Minc/softwaredevelopment/minc2.0
file format reference,” Oct 2012. https://en.wikibooks.org/wiki/MINC/
SoftwareDevelopment/MINC2.0_File_Format_Reference [Accessed; posted 24
January 2019].
[25] O. S. Pianykh, Digital Imaging and Communications in Medicine: A Practical In-
troduction and Survival Guide. Springer Publishing Company, Incorporated, 1 ed.,
2008.
27
[26] J. R. Koza, F. H. Bennett, D. Andre, and M. A. Keane, Automated Design of Both
the Topology and Sizing of Analog Electrical Circuits Using Genetic Programming,
pp. 151–170. Dordrecht: Springer Netherlands, 1996. https://doi.org/10.1007/
978-94-009-0279-4_9.
[27] M. Campbell, A. J. Hoane, Jr., and F.-h. Hsu, “Deep blue,” Artif. Intell., vol. 134,
pp. 57–83, Jan. 2002.
[28] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot,
L. Sifre, D. Kumaran, T. Graepel, T. Lillicrap, K. Simonyan, and D. Hassabis, “Mas-
tering chess and shogi by self-play with a general reinforcement learning algorithm,”
2017.
[29] S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach (3rd Edition).
Pearson, 2009.
[30] M. Mohri, A. Rostamizadeh, and A. Talwalkar, Foundations of Machine Learning
(Adaptive Computation and Machine Learning series). The MIT Press, 2012.
[31] D. H. Wolpert, “The lack of a priori distinctions between learning algorithms,” Neural
Computation, vol. 8, pp. 1341–1390, Oct 1996.
[32] C. Cortes and V. Vapnik, “Support-vector networks,” Mach. Learn., vol. 20, pp. 273–
297, Sept. 1995.
[33] V. N. Vapnik, The Nature of Statistical Learning Theory. Berlin, Heidelberg:
Springer-Verlag, 1995.
[34] O. Chapelle, P. Haffner, and V. N. Vapnik, “Support vector machines for histogram-
based image classification,” Trans. Neur. Netw., vol. 10, pp. 1055–1064, Sept. 1999.
[35] H. Drucker, D. Wu, and V. N. Vapnik, “Support vector machines for spam catego-
rization,” Trans. Neur. Netw., vol. 10, pp. 1048–1054, Sept. 1999.
[36] S.-K. Woo, C.-B. Park, and S.-W. Lee, “Protein secondary structure prediction using
sequence profile and conserved domain profile,” in Advances in Intelligent Computing
(D.-S. Huang, X.-P. Zhang, and G.-B. Huang, eds.), (Berlin, Heidelberg), pp. 1–10,
Springer Berlin Heidelberg, 2005.
[37] N. Cristianini and J. Shawe-Taylor, An Introduction to Support Vector Machines:
And Other Kernel-based Learning Methods. New York, NY, USA: Cambridge Uni-
versity Press, 2000.
[38] R. Kumar, A. Kulkarni, V. K. Jayaraman, and B. D. Kulkarni, “Symbolization as-
sisted svm classifier for noisy data,” Pattern Recogn. Lett., vol. 25, pp. 495–504, Mar.
2004.
[39] S. Mukkamala, A. H. Sung, and A. Abraham, “Intrusion detection using an ensemble
of intelligent paradigms,” J. Netw. Comput. Appl., vol. 28, pp. 167–182, Apr. 2005.
[40] R. K. Prasoona, A. Jyoti, Y. Mukesh, S. Nishant, N. S. Anuraj, and J. Shobha, “Op-
timization of gaussian kernel function in support vector machine aided qsar studies
of c-aryl glucoside sglt2 inhibitors,” Interdisciplinary Sciences: Computational Life
Sciences, vol. 5, pp. 45–52, Mar 2013.
28
[41] O. Z. Maimon and L. Rokach, Data Mining with Decision Trees: Theroy and Ap-
plications (Machine Perception and Artificial Intelligence). World Scientific Pub Co
Inc, 2007.
[44] E. Cantú-Paz and C. Kamath, “Using evolutionary algorithms to induce oblique deci-
sion trees,” in Proceedings of the 2Nd Annual Conference on Genetic and Evolution-
ary Computation, GECCO’00, (San Francisco, CA, USA), pp. 1053–1060, Morgan
Kaufmann Publishers Inc., 2000.
[45] Q. Ding, Q. Ding, and W. Perrizo, “Decision tree classification of spatial data streams
using peano count trees,” in Proceedings of the 2002 ACM Symposium on Applied
Computing, SAC ’02, (New York, NY, USA), pp. 413–417, ACM, 2002.
[48] Z.-H. Zhou, Ensemble Methods: Foundations and Algorithms. Chapman &
Hall/CRC, 1st ed., 2012.
[50] T. K. Ho, “Random decision forest,” pp. 14–16, Proceedings of the 3rd International
Conference on Document Analysis and Recognition, Montreal, Aug. 1995.
[51] Y. Bengio, “Learning deep architectures for AI,” Foundations and Trends
R in Ma-
chine Learning, vol. 2, no. 1, pp. 1–127, 2009.
[54] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,”
CoRR, vol. abs/1512.03385, 2015.
29
[55] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale
image recognition,” CoRR, vol. abs/1409.1556, 2014. http://arxiv.org/abs/1409.
1556.
[60] D. Cid, M. Y., and H., “Texture-based graph model of the lungs for drug resistance
detection, tuberculosis type classification, and severity scoring: Participation in Im-
ageCLEF 2018 tuberculosis task,” CLEF2018 Working Notes, September 10-14 2018.
CEUR-WS.org http://ceur-ws.org.
30